ANN-modell för att bestämma renoveringsbehov av vattenledningar -- Utvärdering av viktiga attribut med tillämpning för Umeå kommun
(2020) In TVVR20/5010 VVRM05 20201Division of Water Resources Engineering
- Abstract (Swedish)
- Sveriges vattenledningsnät kräver kontinuerliga och stora in-
vesteringar och måste underhållas på ett effektivt sätt; syftet med
den här studien var därför att utröna vilka ledningsattribut som
är viktigast för att i en ANN-modell identifiera ledningar med
hög risk för läckage. Detta gjordes genom att först använda attri-
buturvalsmetoderna ReliefF och Recursive Feature Elimination
(RFE) tillsammans med Random Forest Classification (RFC) och
Multinomial logistisk regression (MLR) för att skapa urval av attri-
buten. Dessa grupper av attribut användes sedan i ANN-modellen,
och modellens prestation med respektive urvalsgrupp jämfördes.
ReliefF och RFE med MLR lyckades rangordna attributen, medan
RFE med RFC endast kunde särskilja... (More) - Sveriges vattenledningsnät kräver kontinuerliga och stora in-
vesteringar och måste underhållas på ett effektivt sätt; syftet med
den här studien var därför att utröna vilka ledningsattribut som
är viktigast för att i en ANN-modell identifiera ledningar med
hög risk för läckage. Detta gjordes genom att först använda attri-
buturvalsmetoderna ReliefF och Recursive Feature Elimination
(RFE) tillsammans med Random Forest Classification (RFC) och
Multinomial logistisk regression (MLR) för att skapa urval av attri-
buten. Dessa grupper av attribut användes sedan i ANN-modellen,
och modellens prestation med respektive urvalsgrupp jämfördes.
ReliefF och RFE med MLR lyckades rangordna attributen, medan
RFE med RFC endast kunde särskilja tre attribut som mindre vik-
tiga – dessa rangordningar innebar dock inte att attributen faktiskt
var viktiga – de behövde testas i ANN-modellen först. Studien
visade att antalet attribut kunde begränsas markant: när 10 utvalda
attribut användes uppnåddes en noggrannhet på 0,79, att jämföra
med alla tillgängliga attribut (19 stycken) då en noggranhet på
0,80 erhölls. Effekten av att endast inkludera attribut som är lätta
att anskaffa och som är jämförbara mellan orter undersöktes också
och en modellnoggrannhet på 0,75 uppnåddes då. (Less) - Abstract
- Sweden’s water pipe network demands continuous and large
investments and must be maintained in an effective way; the aim
with this study was therefore to investigate which pipe features are
most important to identify pipes with a high risk of leakage in an
ANN-model. This was done by first utilizing the feature selection
methods ReliefF and Recursive Feature Elimination (RFE) with
Random Forest Classification (RFC) and Multinomial Logistic
Regression (MLR) to identify subsets of features seemingly
important to evaluate the risk of leakage. The ANN-model were
then run with these subsets and the difference in accuracy between
subsets were compared. ReliefF and RFE with MLR succeeded
in ranking features, while RFE with RFC only... (More) - Sweden’s water pipe network demands continuous and large
investments and must be maintained in an effective way; the aim
with this study was therefore to investigate which pipe features are
most important to identify pipes with a high risk of leakage in an
ANN-model. This was done by first utilizing the feature selection
methods ReliefF and Recursive Feature Elimination (RFE) with
Random Forest Classification (RFC) and Multinomial Logistic
Regression (MLR) to identify subsets of features seemingly
important to evaluate the risk of leakage. The ANN-model were
then run with these subsets and the difference in accuracy between
subsets were compared. ReliefF and RFE with MLR succeeded
in ranking features, while RFE with RFC only separated three
features as less important—these rankings, though, had to be
tested with the ANN-model to see if the features actually were
important. The study found that the amount of features could be
reduced distinctly: with 10 important features, an accuracy of 0.79
were achieved, to compare with 0.80 when all the 19 available
features were utilized in the model. The effect of only including
features that are easy to obtain and are similar between cities was
also studied, and a model accuracy of 0.75 were obtained with
these attributes. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9021828
- author
- Nilsson, Didrik LU
- supervisor
-
- Johanna Sörensen LU
- Erik Nilsson LU
- organization
- course
- VVRM05 20201
- year
- 2020
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- underhåll, förnyelse, strategiskt, attributurval, maskininlärning
- publication/series
- TVVR20/5010
- report number
- 20/5010
- ISSN
- 1101-9824
- language
- Swedish
- additional info
- Examiner: Magnus Larson
- id
- 9021828
- date added to LUP
- 2020-06-24 10:29:49
- date last changed
- 2020-06-24 10:29:49
@misc{9021828, abstract = {{Sweden’s water pipe network demands continuous and large investments and must be maintained in an effective way; the aim with this study was therefore to investigate which pipe features are most important to identify pipes with a high risk of leakage in an ANN-model. This was done by first utilizing the feature selection methods ReliefF and Recursive Feature Elimination (RFE) with Random Forest Classification (RFC) and Multinomial Logistic Regression (MLR) to identify subsets of features seemingly important to evaluate the risk of leakage. The ANN-model were then run with these subsets and the difference in accuracy between subsets were compared. ReliefF and RFE with MLR succeeded in ranking features, while RFE with RFC only separated three features as less important—these rankings, though, had to be tested with the ANN-model to see if the features actually were important. The study found that the amount of features could be reduced distinctly: with 10 important features, an accuracy of 0.79 were achieved, to compare with 0.80 when all the 19 available features were utilized in the model. The effect of only including features that are easy to obtain and are similar between cities was also studied, and a model accuracy of 0.75 were obtained with these attributes.}}, author = {{Nilsson, Didrik}}, issn = {{1101-9824}}, language = {{swe}}, note = {{Student Paper}}, series = {{TVVR20/5010}}, title = {{ANN-modell för att bestämma renoveringsbehov av vattenledningar -- Utvärdering av viktiga attribut med tillämpning för Umeå kommun}}, year = {{2020}}, }